A dismissal decided by the algorithm is void: from 23 October in Italy

The part that really changes the job is one line. Article 41 states that final decisions on the employment relationship cannot be based solely on automated processing, that the worker is entitled, on request, to an understandable explanation of the decision, and at paragraph 4 that decisions taken in breach of these rules are void. Not «voidable on appeal»: void. A dismissal, a non-renewal, a change of duties decided by the system alone produce no effect, and that holds regardless of how good the model is and how large the administrative fine might be.

It is worth pausing on the word «solely», because everything turns on it and because it is the same word as in Article 22 GDPR, with ten years of argument behind it. An HR manager who receives a ranking generated by the system, has no way of contesting it, does not know the criteria and signs off what the system proposes, is not human intervention: that is a signature. The intervention that takes the decision outside the rule is by someone with the authority and the information to change it, and who sometimes actually changes it.

• It applies to FINAL decisions on the relationship: hiring, modification, termination. An automated filter upstream of a real interview is a different thing, but it still has to be declared.
• The explanation must be given on request and must be understandable: «the system assigned a score of 42» is not an explanation, it is a number.
• The notice under Article 1-bis of decree 152/1997 must be given BEFORE the processing starts, and covers decision-making and monitoring systems too.
• AI systems affecting work organisation, production rhythms or how work is performed go into the risk assessment, not into a separate annex.
• Workers need information and training on the specific risks of those systems and on the measures adopted.

The Italian picture adds to what the data protection authority had already said. In a decision of 14 May 2026 the authority addressed artificial intelligence systems in the workplace, in particular tools that analyse workers' emotional or psychological aspects, with a principle that is simple to state and uncomfortable to apply: the employer must not come to know, even indirectly, information about the emotional or psychological sphere of those who work for them. A system measuring tone of voice on a customer service call, or posture in front of a camera, falls inside it even when the vendor calls it «quality analysis».

For a DPO there are three things to ask the client, and the first is often disarming. Which systems already in the building produce a score, a ranking or an alert about a person? Almost always the first answer is «none», and almost always what then surfaces is the shift-scheduling software, the warehouse productivity tool, the CV filter in the recruitment portal. None of these was bought as «artificial intelligence», and that is exactly why they are not in the register.

The other two are quicker. Who signs the final decision, and do they genuinely have the power to depart from the suggestion? And what is said to the worker who asks why. It is worth writing that answer before anyone asks for it: a text prepared calmly, with the criteria and the weight of each, is a compliance document; the same text written in a hurry after a complaint is a justification, and it shows.